Separating Synchronization Concerns with Frameworks and Generative Programming

Separating Synchronization Concerns with Frameworks and Generative Programming Scott D. Fleming, R. E. K. Stirewalt, Laura K. Dillon, and Beata Sarna-Starosta Software Engineering and Network Systems Laboratory Department of Computer Science and Engineering Michigan State University East Lansing, Michigan USA 48824 fsdf,stire,ldillon,[email protected] Abstract Thread synchronization in object-oriented systems is difficult to implement, in part, because traditional synchronization mechanisms are low level and, in part, because synchronization code is cross cutting and usually interleaved with functional code. In prior work, we addressed the first of these problems with a compositional model of mutual exclusion called Szumo, that is customized to the needs of strictly exclusive systems. This paper extends our prior work to address the second problem. We present SzumoC++, an approach to building strictly exclusive systems in C++ that affords a high degree of abstraction and separation of concerns. Development in SzumoC++ proceeds from an explicit design model, which is assumed to have been validated. Our approach uses object-oriented frameworks and a custom specification language to support the development of separate functional and synchronization modules which retain a high degree of design transparency. By virtue of this transparency, and with the aid of generative programming techniques, these separately implemented modules may be composed in a consistent manner. By virtue of being based on a Szumo design, we expect the resulting systems to be robust under change. 1 Introduction Thread synchronization in object-oriented (OO) systems is a global and cross-cutting concern, which is difficult to implement and is resistant to clean separation from the primary functional logic of programs. While much progress has been made in the development of tools and frameworks for separating the synchronization from the functional logic of a program (e.g., [18]), synchronization concerns are difficult to modularize in such a manner that local changes to the functional logic engender commensurately local changes to the synchronization logic. Consequently, accommodating a local change or extension to the functional logic could require a redesign of the program’s synchronization logic. This problem is especially vexing during software maintenance and is detrimental to the construction of open and adaptable systems. Our work aims to support the development and long-term maintenance of multi-threaded OO systems that are robust under change. We believe this goal can be achieved if the methods used to separate synchronization from functional logic are based on clean compositional design models, and if the separated synchronization and functional modules exhibit a high degree of transparency with respect to the design. This design transparency enables both the automated composition of and a form of consistency checking between the functional and synchronization modules. Previously, we developed a compositional model of synchronization called the Synchronization Units Model (Szumo) [2, 25]. Szumo trades generality for compositionality by focusing on the category of strictly exclusive systems, in which multiple threads compete for exclusive access to dynamically changing sets of shared resources. This narrowing of focus was motivated by the observation that many applications fit well in this category and that, in such cases, we can exploit a clean, compositional model of design and verifica- 1 tion [22, 9]. Examples of strictly exclusive systems include extensible web servers and interactive applications with complex graphical user interfaces. Our previous implementation of Szumo involved an extension of an existing programming language (Eif- fel) with new constructs that directly express the features of a Szumo design [1, 25]. Using this extended language, we were able to show how Szumo supports modular descriptions of the synchronization concern, raising the level of abstraction at which synchronization concerns are expressed, and automating implementation and enforcement of these concerns. A case study involving maintenance and evolution of a web server, based on the architectural design of Apache, demonstrated that a well-designed Szumo application could be safely modified and extended without requiring redesign of the original synchronization concern [3]. How- ever, this implementation of Szumo lacked facilities for separating synchronization logic and functional logic to the extent achievable in approaches such as [4, 8, 18]. Upon reflection, we realized that an implementation of Szumo based solely on a language extension could not easily accommodate such facilities, especially in a modern production language such as C++. This paper describes a new approach to programming Szumo applications that overcomes these deficien- cies. The approach builds on explicit models of a Szumo design and a new programming system, called SzumoC++. A prototype implementation of our programming system can be found at: http://www.cse.msu.edu/sens/szumo/szumoc++.tar.gz. SzumoC++ uses a novel combination of plug-compatible OO frameworks and generative programming to impose upon the separated functional and synchronization modules a level of design transparency that enables the safe and automatic composition of concerns.The result simplifies programming of the synchronization concern and supports a separation of functional and synchronization concerns that inherits Szumo’s support for robustness under change. In the remainder of the paper, we first give a background on the Szumo model and a sample Szumo design to use in a running example (Section 2). We then provide an overview of our approach (Section 3), deferring the details of its many components to later sections. We conclude with a comparison with related work (Section 8), and a discussion of lessons learned and possible directions for the future work (Section 9). 2 Szumo Irrespective of how it is implemented (i.e., as a language extension or as part of a programming system), Szumo is a model of synchronization among entities in a strictly exclusive program. In this model, threads contend with one another for exclusive access to sets of synchronization units. Intuitively, the synchronization units in an application define indivisible units of sharing—at any time during execution, a thread is allowed to access either all of the objects contained in a synchronization unit or none of the objects contained in that unit [1, 25]. In lieu of code that invokes OS-level synchronization primitives, a synchronization unit declares one or more synchronization constraints, which specify the conditions under which the unit—acting as a client—needs exclusive access to units—acting as the client’s direct suppliers—that the client holds refer- ences to. At run time, threads negotiate for exclusive access to suppliers in accordance with the constraints declared by any clients that they are executing. The declarative nature of synchronization constraints and this automated process of negotiation are key to Szumo’s support for extension and maintenance [3]. As a background for subsequent sections, we now describe the details of a Szumo design, and provide an overview of the negotiation process. 2.1 Design Models A key component of our approach is the use of an explicit model of a program’s design, which is used both as a reference during implementation and also during the composition of separated concerns. Szumo design 2 <<root_unit>> left <<unit>> Philosopher Philosopher right Fork before eat() eating: bool <<unit>> ! eating eating void think() sync_constr =<<unit>> "eating ==> left" void eat() sync_constr = "eating ==> right" after eat() (a) (b) Figure 1: Diagrams depicting the unit-class and synchronization-state models of a Szumo design models are expressed using a variant of the class and state diagram notations of UML 2.0 [21]. These models impose a modular structure upon and record key design decisions related to the synchronization concern. An important design decision involves the granularity of sharing among the threads. Szumo allows application developers to choose the granularity of sharing by means of synchronization units, which are containers of (one or more) program objects. Synchronization units themselves resemble objects in that they may en- capsulate state, provide operations, and reference other synchronization units whose operations they use to implement their own. When a program object is created, it is deployed to exactly one synchronization unit, where it remains throughout its lifetime. A thread that holds exclusive access to a synchronization unit, holds exclusive access to all objects contained within this unit. Synchronization units are identified with instances of types called unit classes. In addition to standard operations, a unit class C may declare unit variables, condition variables, and synchronization constraints. Unit variables reference synchronization units that serve as direct suppliers to units of type C. Condition variables are boolean-valued variables in a unit class. The synchronization constraints specify the suppliers to which a client unit requires exclusive access based on the values of the unit’s condition variables and unit variables. We say that a client unit c entails a direct supplier s when c requires exclusive access to s, and refer to the set of all direct suppliers that c entails at any given time as c’s entailment. To illustrate these ideas, Fig. 1 depicts the models used

Load more